H3 API
MiniMax H3 is a 33B-parameter open-weights diffusion transformer that unifies text, image, video, and audio understanding with native video and stereo audio generation.
MiniMax H3 is an omni-modal generative model from MiniMax, released as open weights on August 3, 2026 under the MiniMax H3 Community License Agreement. It is built around the H3-Base core model, which combines a text encoder, a visual VAE, an audio VAE, and an omni transformer; RoPE is used to capture spatial and temporal relationships. The system also includes H3-Context-IR, a preprocessing and orchestration layer that is not part of the open-source release and is available only via API. A new high-compression tokenizer expands the effective sequence length fourfold and helps cut training and inference cost.
The model accepts text, image, video, and audio inputs and understands the relationships between them. From a single model it generates video and audio together, at 480p, 768p, or native 2K resolution, at 24FPS and with lengths from 5 to 15 seconds. It supports first-frame and last-frame control, image-to-video conversion, video-to-video motion transfer, voice timbre reproduction from audio references, and native stereo audio covering voice, sound effects, and music. It follows complex multimodal instructions in natural language and handles editing tasks such as lip sync to new dialogue and partial reuse of background music, while rendering on-screen text and brand expression accurately.
These abilities suit advertising, branding, e-commerce, product design, UI/UX, game content production, short-form video production with reference material, and instruction-based transitions. MiniMax describes the model as top-tier among open-source video generation models. Physical-world consistency remains a challenge, however: visual plausibility does not guarantee physical consistency, multimodal input does not always lead to effective multimodal reasoning, and reported failures include incorrect evidence grounding, incomplete event realization, and violations of contact mechanics and object geometry.
Updated:
MODEL ID
hailuo-3- Input
- text / image / video
- Output
- video
Highlights
- 33B-parameter diffusion transformer
- Unified multimodal input and output
- Native 2K video with stereo audio
- 24FPS, 5 to 15 second clips
- Open weights, community license
- H3-Context-IR available via API
Providers
Compare API prices across providers. Select a provider to view all charges and conditions.
Video specifications
- Aspect ratios
- 21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16
- Audio · Output
- Yes
- Duration · Minimum
- 5 seconds
- Duration · Maximum
- 15 seconds
- Duration · Step
- 1 seconds
- Resolutions
- 2K
API reference
Model-specific protocol details are not published. Check the API reference for integration requirements.
API reference →Frequently asked questions
What is H3?
MiniMax H3 is an omni-modal generative model from MiniMax, released as open weights on August 3, 2026 under the MiniMax H3 Community License Agreement. It is built around the H3-Base core model, which combines a text encoder, a visual VAE, an audio VAE, and an omni transformer; RoPE is used to capture spatial and temporal relationships. The system also includes H3-Context-IR, a preprocessing and orchestration layer that is not part of the open-source release and is available only via API. A new high-compression tokenizer expands the effective sequence length fourfold and helps cut training and inference cost. The model accepts text, image, video, and audio inputs and understands the relationships between them. From a single model it generates video and audio together, at 480p, 768p, or native 2K resolution, at 24FPS and with lengths from 5 to 15 seconds. It supports first-frame and last-frame control, image-to-video conversion, video-to-video motion transfer, voice timbre reproduction from audio references, and native stereo audio covering voice, sound effects, and music. It follows complex multimodal instructions in natural language and handles editing tasks such as lip sync to new dialogue and partial reuse of background music, while rendering on-screen text and brand expression accurately. These abilities suit advertising, branding, e-commerce, product design, UI/UX, game content production, short-form video production with reference material, and instruction-based transitions. MiniMax describes the model as top-tier among open-source video generation models. Physical-world consistency remains a challenge, however: visual plausibility does not guarantee physical consistency, multimodal input does not always lead to effective multimodal reasoning, and reported failures include incorrect evidence grounding, incomplete event realization, and violations of contact mechanics and object geometry.
How is usage billed?
Your bill depends on the provider you use, its prices and your usage. Longer inputs or other conditions may have different prices. Billed based on the actual generation cost. See your bill for the final amount.
What inputs does MiniMax H3 accept?
It accepts text, image, video, and audio inputs and understands the relationships between them, following complex multimodal instructions described in natural language.
What does MiniMax H3 output?
It generates video and audio from a single model. Output resolutions are 480p, 768p, and 2K, at 24FPS with lengths from 5 to 15 seconds, and the audio is native stereo covering voice, sound effects, and music.
Which video tasks can MiniMax H3 handle?
It performs image-to-video conversion and video-to-video motion transfer, supports first-frame and last-frame control, and handles editing tasks such as lip sync to new dialogue and partial reuse of background music.
How does MiniMax H3 relate to MiniMax H3 Max?
MiniMax H3 Max is a variant specialized for fast, high-volume 480p or 768p generation, while MiniMax H3 is the 33B-parameter omni-modal model released as open weights.
What are the known limitations?
Physical-world consistency remains a challenge, multimodal input does not always lead to effective multimodal reasoning, and reported failures include incorrect evidence grounding, incomplete event realization, and violations of contact mechanics and object geometry.
